Google DeepMind's Gemini Deep Think Advances AI-Assisted Mathematics and Physics Research

Google DeepMind has published two research papers demonstrating how its Gemini Deep Think mode is being deployed to tackle frontier research problems in mathematics, physics, and computer science—moving well beyond academic competitions.

From Competition to Research#

Earlier this year, Gemini Deep Think achieved Gold-medal standard at the International Mathematics Olympiad (IMO) and made similar progress at the International Collegiate Programming Contest. Building on that foundation, the model has now advanced into professional research workflows, where it collaborates with expert mathematicians, physicists, and computer scientists on open-ended, complex challenges.

Solving Research-Level Mathematics#

The first paper introduces Aletheia, a math research agent powered by Gemini Deep Think. Unlike IMO-level problems, research mathematics requires synthesizing vast literature and advanced techniques—a domain where foundation models often struggle with data scarcity and hallucinations.

Aletheia incorporates several key capabilities:

• Iterative verification: A natural language verifier identifies flaws in candidate solutions, enabling cycles of generation and refinement
• Graceful failure recognition: The agent can admit when it cannot solve a problem, improving efficiency for human researchers
• Web search integration: Access to Google Search prevents spurious citations and computational errors when synthesizing published literature

Performance metrics show Aletheia scoring up to 90% on the IMO-ProofBench Advanced test, with scaling improvements continuing into PhD-level mathematics.

Real-World Research Impact#

The work has already contributed to published mathematics through three tiers of collaboration:

• Autonomous research: A peer-reviewed paper on eigenweights in arithmetic geometry, generated entirely by AI
• Human-AI collaboration: Joint work proving bounds on systems of interacting particles
• Semi-autonomous evaluation: Analysis of 700 open problems, including autonomous solutions to four previously unsolved questions from the Erdős Conjectures database